Home
last modified time | relevance | path

Searched refs:vk02c1 (Results 1 – 4 of 4) sorted by relevance

/external/XNNPACK/src/f32-conv-hwc2chw/
D3x3s2p1c3x4-neonfma-2x2.c277 const float32x4_t vk02c1 = vld1q_f32(w + 88); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2() local
279 vo0x0 = vfmaq_laneq_f32(vo0x0, vk02c1, vi0x2, 0); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2()
280 vo1x0 = vfmaq_laneq_f32(vo1x0, vk02c1, vi2x2, 0); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2()
281 vo0x1 = vfmaq_laneq_f32(vo0x1, vk02c1, vi0x3, 2); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2()
282 vo1x1 = vfmaq_laneq_f32(vo1x1, vk02c1, vi2x3, 2); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2()
553 const float32x4_t vk02c1 = vld1q_f32(w + 88); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2() local
555 vo0x0 = vfmaq_laneq_f32(vo0x0, vk02c1, vi0x2, 0); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2()
556 vo1x0 = vfmaq_laneq_f32(vo1x0, vk02c1, vi2x2, 0); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2()
D3x3s2p1c3x4-neon-2x2.c277 const float32x4_t vk02c1 = vld1q_f32(w + 88); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2() local
279 vo0x0 = vmlaq_lane_f32(vo0x0, vk02c1, vget_low_f32(vi0x2), 0); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
280 vo1x0 = vmlaq_lane_f32(vo1x0, vk02c1, vget_low_f32(vi2x2), 0); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
281 vo0x1 = vmlaq_lane_f32(vo0x1, vk02c1, vget_high_f32(vi0x3), 0); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
282 vo1x1 = vmlaq_lane_f32(vo1x1, vk02c1, vget_high_f32(vi2x3), 0); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
551 const float32x4_t vk02c1 = vld1q_f32(w + 88); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2() local
553 vo0x0 = vmlaq_lane_f32(vo0x0, vk02c1, vget_low_f32(vi0x2), 0); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
554 vo1x0 = vmlaq_lane_f32(vo1x0, vk02c1, vget_low_f32(vi2x2), 0); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
D3x3s2p1c3x4-wasmsimd-2x2.c278 const v128_t vk02c1 = wasm_v128_load(w + 88); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2() local
280 …vo0x0 = wasm_f32x4_add(vo0x0, wasm_f32x4_mul(vk02c1, wasm_v32x4_shuffle(vi0x2, vi0x2, 0, 0, 0, 0))… in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
281 …vo1x0 = wasm_f32x4_add(vo1x0, wasm_f32x4_mul(vk02c1, wasm_v32x4_shuffle(vi2x2, vi2x2, 0, 0, 0, 0))… in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
282 …vo0x1 = wasm_f32x4_add(vo0x1, wasm_f32x4_mul(vk02c1, wasm_v32x4_shuffle(vi0x3, vi0x3, 2, 2, 2, 2))… in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
283 …vo1x1 = wasm_f32x4_add(vo1x1, wasm_f32x4_mul(vk02c1, wasm_v32x4_shuffle(vi2x3, vi2x3, 2, 2, 2, 2))… in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
553 const v128_t vk02c1 = wasm_v128_load(w + 88); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2() local
555 …vo0x0 = wasm_f32x4_add(vo0x0, wasm_f32x4_mul(vk02c1, wasm_v32x4_shuffle(vi0x2, vi0x2, 0, 0, 0, 0))… in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
556 …vo1x0 = wasm_f32x4_add(vo1x0, wasm_f32x4_mul(vk02c1, wasm_v32x4_shuffle(vi2x2, vi2x2, 0, 0, 0, 0))… in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
D3x3s2p1c3x4-sse-2x2.c277 const __m128 vk02c1 = _mm_load_ps(w + 88); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_2x2() local
279 …vo0x0 = _mm_add_ps(vo0x0, _mm_mul_ps(vk02c1, _mm_shuffle_ps(vi0x2, vi0x2, _MM_SHUFFLE(0, 0, 0, 0))… in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_2x2()
280 …vo1x0 = _mm_add_ps(vo1x0, _mm_mul_ps(vk02c1, _mm_shuffle_ps(vi2x2, vi2x2, _MM_SHUFFLE(0, 0, 0, 0))… in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_2x2()
281 …vo0x1 = _mm_add_ps(vo0x1, _mm_mul_ps(vk02c1, _mm_shuffle_ps(vi0x3, vi0x3, _MM_SHUFFLE(2, 2, 2, 2))… in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_2x2()
282 …vo1x1 = _mm_add_ps(vo1x1, _mm_mul_ps(vk02c1, _mm_shuffle_ps(vi2x3, vi2x3, _MM_SHUFFLE(2, 2, 2, 2))… in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_2x2()
553 const __m128 vk02c1 = _mm_load_ps(w + 88); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_2x2() local
555 …vo0x0 = _mm_add_ps(vo0x0, _mm_mul_ps(vk02c1, _mm_shuffle_ps(vi0x2, vi0x2, _MM_SHUFFLE(0, 0, 0, 0))… in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_2x2()
556 …vo1x0 = _mm_add_ps(vo1x0, _mm_mul_ps(vk02c1, _mm_shuffle_ps(vi2x2, vi2x2, _MM_SHUFFLE(0, 0, 0, 0))… in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_2x2()